JMIR Publications

ABSTRACT

AI systems in preventive health fail not primarily because of poor prediction, but because they are not designed to support sustained behavioural processes. Many AI systems are designed as episodic prediction tools, yet preventive care depends on sustained behavioural change over time. As a result, systems that perform well in controlled settings may fail to produce meaningful improvements in real-world outcomes. This Perspective argues that this limitation reflects a structural misalignment between prevailing AI paradigms and the temporal dynamics of health behaviour. We propose a shift toward interaction-centric AI, in which system outputs are embedded within continuous loops linking insight, action, feedback, and adaptation. We present a behavioral–informatics framework that integrates data, knowledge, reasoning, interaction, behaviour, and governance, highlighting the interdependencies required to support longitudinal engagement. We show that existing approaches, including behavioural intervention frameworks and human AI interaction models, address these elements in isolation but do not resolve their integration. We further argue that current evaluation strategies, centred on predictive accuracy and short-term engagement, are insufficient for systems whose effects emerge over time. Instead, evaluation should prioritise behavioural trajectories, sustained adherence, and long-term outcomes. Interaction-centric systems introduce new challenges, including risks of behavioural influence, inequity, and dependency, raising important clinical and policy questions. Addressing these challenges is essential if AI is to move beyond prediction and contribute meaningfully to preventive health. This perspective situates interaction-centric AI within the broader context of digital behaviour change interventions.